When carrying out a statistical analysis of extremes, if left undetected spurious observations can greatly affect 50-year return period wind speed estimates. Methods for identifying such observations require attention, particularly for the current work where a subset of maxima is sought and errors are known to exist within the ISD. Quality control measures can also aid in determining if annual mean wind speeds are consistent over the entire data record or if considerable discontinuities exist. Identification of a shift may indicate changes of instrument location, height or local surroundings.
Throughout the operational lifetime of a synoptic weather station, it is not uncommon for the height or location of the anemometer to change, or for instrumentation to be upgraded. Quite often a meteorological agency will upgrade all instrumentation for a given date, al- though in practice it may be several months before the upgrades are operational at every location. Each of the possible changes will have a specific effect on the measured wind speed. A change of anemometer height will be most apparent from a shift of the mean wind speed for all directions, however, a change of anemometer location can be much more complex. If the old and new site have very similar exposures for all directions, a change may not be detectable unless otherwise documented. Conversely, a new location where the exposure differs directionally from the prior location will experience changes of mean wind speed and gustiness in the affected directions. The gustiness at a site can be represented by the gust factor, the ratio of the gust wind speed to mean wind speed. Lastly, a change of the anemometer or chart recorder should not be apparent from the mean wind speed records, as any changes to the response length or gust averaging time should be filtered out over a suf- ficient averaging period. The change will most likely affect measurements of the gustiness.
Documentation is occasionally available from meteorological agencies identifying dates of location or instrumentation changes, if unavailable, it is important to identify these changes to at least be aware that they exist.
At present, only two 50-year return period wind speed studies identify the methods cho- sen to pre-process meteorological data. Sacr´e et al. (2007) implement a detection method called PRODIGE which is described by Caussinus and Mestre (2004) and used by M´et´eo- France. The PRODIGE algorithm is applied to annual mean data from stations assumed to be influenced by the same climatic conditions. Each series is assumed to be a combination of a climate effect, station effect and random white noise (Caussinus and Mestre, 2004). In performing the analysis across multiple stations, the climatic effect should be spatially redundant, thus allowing differences due to station effects to be identified. A penalised log-likelihood procedure of Caussinus and Lyazrhi (1997) is used to detect change-points and outliers, and least-square estimates of the climate and station effects are used to correct the data. To vastly reduce the number of hypotheses and computational time, a prelimi- nary stage consisting of pairwise comparisons of the station record with those from neigh- bouring locations is required. These difference series, in conjunction with the penalised log-likelihood procedure and manual synthesis, are used for pre-selection of change-points and outliers in monthly or annual mean data. A procedure which can be automated with- out requiring a pre-selection stage is preferred for the number of stations considered here. In addition, the resulting ‘corrected’ data may not be appropriate for wind observations particularly those exhibiting significant directional variation. For an anemometer sited in relatively open terrain, a change of height will likely have an isotropic influence on the wind speeds, in this situation a single station effect will be appropriate for all wind speeds measured at the location. However, if the location of an anemometer has changed, then differences in surface roughness may only occur for certain azimuths, therefore, the true station effect may exhibit anisotropy. Caussinus and Mestre (2004) note the PRODIGE
model requires better detection of gradual changes and of breaks when the shift of the mean is less than the standard deviation.
Burton and Allsop (2009a) pre-process wind speed data in an attempt to identify individual observations for removal. Mean wind speeds greater than 20 m/s and three times greater than both adjacent mean hourly observations are classified as errors or thunderstorms, both of which are excluded from a synoptic climate analysis. For a number of regions in Europe of interest in the current work, the 50-year return period wind speed is less than 27 m/s, as was shown in Figure 2.2. A representative set of annual maxima will likely contain a subset of extremes which are less than 20 m/s, therefore, the maxima contained within the subset are not necessarily validated e.g. an annual maximum of 18 m/s is not considered by the pre-processing scheme. Such a situation is likely to arise, particularly when eval- uating directional extremes where maxima occurring from non-dominant wind directions are, in general, substantially lower than dominant wind directions. A lower threshold of 15 m/s suggested by Burton and Allsop (2009b) is likely more appropriate. Ideally, a method which can be applied to ensure the quality of every hourly wind observation is desired. The data can then be used to accurately derive the parent distribution if desired and, more im- portantly, ensures the validation of maxima regardless of the strength of the wind climate. A combination of the aforementioned quality control measures are required. Quality con- trol measures are divided into two levels for the current work, global or high-level quality control measures and localised or low-level quality control measures. The global quality control measures include physical limits checks based on DeGaetano (1997) and a ho- mogenisation algorithm by Domonkos (2011). Localised quality control measures are based on the wind speed variability checks established by DeGaetano (1997) for hourly surface measurements and are expanded to consider additional information relevant to the current analysis. Additional quality control measures for a range of meteorological param- eters are discussed by Graybeal et al. (2004).